Towards Federated Learning at Scale: System Design

Hubert Eichner
Wolfgang Grieskamp
Dzmitry Huba
Vladimir Ivanov
Chloé M Kiddon
Jakub Konečný
Stefano Mazzocchi
Timon Van Overveldt
David Petrou
Jason Roselander
SysML 2019

Abstract

Federated Learning is a distributed machine learning approach which enables training on a large corpus of data which never needs to leave user devices. We have spent some effort over the last two years building a scalable production system for FL. In this paper, we report about the resulting high-level design, sketching the challenges and the solutions, as well as touching the open problems and future directions.